Impacts of Weather on Short-Term Metro Passenger Flow Forecasting Using a Deep LSTM Neural Network

Round 1

Reviewer 1 Report

This paper dealt with an interesting topic on the dynamics of Metro passenger flow according to weather. Nevertheless, there still exist a room that the manuscript can be improve for readers to understand their experimental design and results better. See the following comments    #1. The expression of "LSTM_NN" is not common, and it may look like a programming code. It can be replace with another expression For example, in the title, you can describe LSTM neural network, and then it can be shorten as ""LSTM-NN".   #2. In Figure 3, the scale of left Y-axis in Figure 3(b) (i.e. 25) should be modified the same as the other two sub-figures (a) and (c) (i.e. 18) for readers to compare the value in an easier way.    #3. In Figure 3, I wonder why the fluctuation of black one (i.e. hourly passenger flow (has rainfall) in Figure 3(c) is serious compared to the other two. The authors have better commnet about that in the main text.   #4. In Figure 5, the connection of arcs is difficult to understand for me. To estimate the hat of Y_t, is it possible to use the value of h_a, which is unknow yet? In addition, in the figure, h_k needs to be corrected to h_t.   #5. In Table 2, the input variables have been provided well. Did you consider all the variables are numerical variables, not categorical variables? If there is any variable that were considered a categorical variable, it needs to be described in Table 2.   #6. I have curiosity to use the 31days as an input variable. For example, the samples with value "January" or "Monday" may have similar passenger flow patterns. But, it is difficult to say the 12 samples with value "day 1" have similar passenger. What is the opinion of the authors?   #7. How did you decide the state-of-the-art methods? The authors simply have cited the related paper of each method. If you cannot provide comparative references that say the four methods generally showed the best performance in this problem (i.e. prediction of passenger flow), it is difficult to say they are the state-of-the-art methods. Moreover, the last three ones are acceptable to be compared with the proposed method. But, the RF is not enough among so many machine learning methods. The authors should provide a few more representative methods such as XGBoost, GradientBoost. The single prediction models such as SVR and Lasso may be omitted.   #8. The performance measures related to "prediction errors" such as RMSE and MAPE are not sufficient to say the effectiveness of prediction. R-squared needs to be provided along with the error-based performance measures. Without R-squared, we cannot compare the regression results if they problem has different scales and variations although MAPE was provided.   Typos: In line #79 : Bo cker -> Bocker

Author Response

Many thanks for your helpful comments.
Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Dear authors,

the paper represents both a highly topical issue for traffic planning and a discussion of the potential of introducing new variables into the analysis of passenger flow forecasting. Methodology and results are presented in a comprehensive and sound manner and especially interesting in regard to combining different existing data sets into reliable forecasting models. Only minor revisions are suggested.

Main comments are concerning writing an English language used. A thorough spellcheck is suggested - some examples below:

• 54: "one weather variables" - singular
• 68: "selected weather condition(.." - missing blank space
• 79: "Bo cker" - Maybe an issue with the pdf conversion of this specific name?
• 85: "contributed individual" - contributed to …

Other than these minor revisions it would interesting to see some of the addressed concepts more elaborated in the context of the study. Pleasant weather conditions are mentioned (176), however as outdoor thermal comfort is a complex topic in itself (i.e. Nikolopoulou, Marialena. (2011). Outdoor thermal comfort. Frontiers in bioscience (Scholar edition). 3. 1552-68. 10.2741/245.) and represents a central aspect of this concept, a more elaborate discussion of the actual situation would help to understand basic weather conditions within the testbed and how this relates to existing concepts.

In the conclusions some general deductions and practical implications would be very interesting. What are the actual implications for future traffic and spatial planning? Smart city concepts are addressed in the introduction section of the article; hence it would be very intriguing to see how these findings can support the actual vision of the smart city.

Travel psychology is alluded to, but actual relevance of understanding socio-psychological aspects of travel behaviour is never discussed in the state-of-the-art section of the paper. This seems to be an important aspect for both interpretation of the results of the model and for actual implications for implementation of derived measures. Seasonal variations (tourism, holidays, etc.) also may have an essential impact on travel behaviour at specific traffic nodes and could support the drawn conclusions.

 

Kind regards

 

Author Response

Many thanks for your helpful comments.
Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The paper is about short-term data-driven analysis and prediction of metro passengers influenced by the weather conditions. The authors have already published on similar themes using other machine learning techniques, probably using same dataset. So naturally, I was looking for "what exactly is the impact of weather", in addition to what is told by the data. In conclusion, there was no surprise that I could find. Nevertheless, prediction was the main idea of the paper! In section 4.5, authors have mentioned that deep LSTM performs better than other techniques. Yes, it has shorter error bars, but what does these values mean (provide y-axis label, also applicable to other graphs)?

Also, I was not really comfortable, with many models numbers. It should be structured and put into proper categories. In fact you do not need to enumerate each and every model. It makes the user confused. Maybe it would be better to provide mean values. Also, how much LSTM is better in percentage. This is how a technique is compared with others!

The other things in paper are fine. English is good (probably still need a final look). Presentation of models and results is good. However, the structure is noisy and can easily be improved.

Author Response

Many thanks for your helpful comments.
Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The comments have been reflected well in the manuscript.